Connect your providers
Configure and test your organisation's model, transcription and voice accounts.
You need an organisation owner or administrator account. Open the selected organisation's Integrations to see available connections.
Configure a working voice stack
- Choose an available model provider and save your organisation's credentials.
- Configure a transcriber and voice provider for voice conversations. An OpenAI configuration can supply the key for all three OpenAI stages; select a compatible installed model in each assistant tab.
- Run Test saved configuration for each connection. A saved key means configured; a successful test means verified at that time. Changes to credentials, endpoints or options require a new test.
- In Build → Assistants, select providers and models in Transcriber, Model and Voice. Choose the recognition language and an available voice.
- Save, then Talk to test the complete conversation.
Provider tests and previews can incur provider usage and Sigsen model credits. Text-only assistants need a model; add speech providers when enabling voice notes or live audio.
Separate platform speech and admin keys
Platform administrators can open a speech provider in Admin → Integrations and save two encrypted credentials:
- Speech API key: used for transcription, speech generation, previews and provider tests.
- Admin API key (optional): used for supported catalog discovery and account reporting. Grant the read permissions those operations need. This key is excluded from assistant runtime credentials. Each key can be replaced or cleared independently; leaving a replacement blank keeps the saved value.
For Deepgram, add the optional Deepgram project ID to discover project-specific voices with the admin key. Without an admin key, project discovery uses the existing speech key. Leave the project ID blank to list public voices without sending a key. Deepgram model metadata documentation explains the two catalogs.
In Configure Deepgram, Voice ID and Model are dropdowns populated from that catalog. Choosing either keeps both fields on the same canonical voice model ID. The test voice and model use the same choices. Save changes to keys or the project ID before refreshing the catalog. Saved values that are no longer listed remain visible until you choose a replacement.
In Configure ElevenLabs, Voice ID and Model are independent dropdowns populated from the account's voices and supported streaming speech models. Save a separate Admin API key with Voices → Read, Models → Access, and User → Access. Speech generation permissions belong on the speech key. Catalog discovery requires the admin key and never falls back to the speech key. Refresh the list after saving changes. ElevenLabs v3/v4 models require a dialogue connection and are excluded from the current TTS connection. ElevenLabs lists the two streaming connections.
ElevenLabs account monitoring shows remaining subscription credits, usage, plan status, reset date and reported overage. These come from the subscription endpoint; the allowance is not a USD cash balance. Enabled accounts refresh automatically every 15 minutes while the local services are running.
Assistant voice settings also list the connected Deepgram, ElevenLabs and Cartesia voices and models. Deepgram keeps the voice and model on the same Aura ID; ElevenLabs and Cartesia selections are independent. Saved values missing from a refreshed list stay visible until replaced. Catalog requests respect the active organization's integration availability and credentials. Platform ElevenLabs catalogs use its explicit admin key; tenant catalogs use that tenant's own key with catalog read permissions.
Cartesia requires its standard speech key for voice discovery and a separate admin key for credit usage reporting. Its model dropdown uses the documented stable Sonic catalog because there is no documented TTS model-list endpoint. Retired models and previews are excluded. Scheduled Cartesia monitoring shows credits consumed by day and month; it does not infer a remaining credit balance or USD spending from consumption.
In the admin Models page, configure Account monitoring and choose the Deepgram or ElevenLabs integration under Reporting credential source. Reporting reads that integration's admin key directly, so rotation takes effect without copying the key. Set the matching Deepgram project UUID for balances and usage. Reporting never falls back to a speech key; missing, removed or deactivated admin credentials stop reporting. Existing separate monitoring keys continue to work, and OpenAI reporting still uses its separate OpenAI Admin API key.
These platform credentials remain private to Sigsen. Other organisations continue using their own provider connections.
Test each stage
| Test | What it establishes |
|---|---|
| Model test | The saved account can generate a short response with the selected model. |
| Voice test or Generate preview | The provider can synthesize the selected voice; a preview can use unsaved editor values. |
| Transcription test | The provider can transcribe your sample audio. |
| Talk | The combined stack runs in a live browser conversation. |
For the integration transcription test, use a mono PCM16 WAV of at most 15 seconds and 2 MiB. An optional reference transcript helps measure recognition errors. Use samples representative of your callers' accents and languages.
When a listed provider fails
Check the account's model access, balance, region and exact model ID. A catalog entry does not prove that your account can run it. Missing or revoked credentials block the request; another tenant's account is never substituted.
Configure fallbacks with working credentials in the same organisation, and review the failed stage in call logs.